Text-to-Speech (TTS)

Text-to-speech (TTS) is the conversion of written text into audible, natural-sounding speech using a trained neural voice model. Modern TTS (ElevenLabs, OpenAI's TTS API, Google Cloud TTS, Amazon Polly, PlayHT) has moved far beyond the robotic concatenative synthesis of the 2000s; today's systems use neural vocoders and diffusion- or transformer-based architectures that model prosody, intonation, breathing, and emotional inflection, producing audio that is frequently indistinguishable from a human recording, complete with natural pauses, filler emphasis, and micro-variations in pitch that make long-form listening feel less fatiguing than older robotic voices. Why it matters for SaaS builders: TTS unlocks accessibility (screen readers, audio versions of articles), audio content products (AI-narrated audiobooks and podcasts, e.g., ElevenReader), IVR/call-center automation, and voice UI for apps. It's a standard building block in customer-support bots, language-learning apps, and video-generation pipelines (voiceover track). Integration is almost always API-based: send text plus a voice ID and get back an audio stream (MP3/WAV) or, for low-latency use cases like voice agents, a real-time audio chunk stream over WebSocket. A concrete worked example — adding "listen to this article" to a blog: (1) on publish, the backend sends the article's plain-text body (with markdown stripped and headings converted to natural pause markers) to the TTS API: `POST /v1/text-to-speech/{voice_id}` with body `{"text": "...", "model_id": "eleven_multilingual_v2", "voice_settings": {"stability": 0.5, "similarity_boost": 0.75}}`; (2) the API returns an MP3 stream, typically generated in roughly real-time-to-real-time-plus-a-few-seconds for short articles; (3) the app stores the file in object storage behind a CDN and embeds an `<audio>` player with playback-speed controls above the article; (4) for long articles that exceed the API's per-request character limit, the backend splits the text into paragraph-level chunks, generates audio for each in parallel, and stitches the resulting MP3 segments together server-side before caching the final file, so it's only generated once per article rather than on every page view. Key parameters: voice selection (stock voices vs. custom cloned voices), stability (consistency vs. expressiveness trade-off — higher stability sounds more monotone but more reliable), speaking rate/speed, and output format/sample rate (44.1kHz MP3 for web playback vs. lower-bitrate formats for bandwidth-constrained mobile). Latency matters enormously for conversational use cases — batch TTS for a blog post can tolerate several seconds of generation time, but providers now offer streaming TTS with sub-300ms time-to-first-audio-byte specifically for real-time voice agents, where the user would otherwise perceive an awkward pause before the AI "starts talking."